YouTube flagged creator Theo's video on Google's Gemini CLI tool and Antigravity CLI 1.0.0 release, triggering monetization limits and a 70% drop in new views as engineers confirmed internal escalation

You know us dude and I at least know you ... there are some things here it's not my place to comment on publicly and I can't comment on or control other people's thoughts or actions.

That said, when you reached out about the monetization issue, both @DeveloperFish and I escalated it immediately. I personally do not believe that the video should have been flagged. But I don't work at YouTube and I don't speak for them. I do think we got it resolved quickly. That's not an excuse or justification and I can't render a disposition. I wish it had never happened. We do the best we can sometimes.

As far as you being persona non grata at Google, not to me you're not.

You're a member of the Google Cloud Startup program, you are a good customer of Google Cloud, and you and I have a good relationship. If you feel that you are facing retaliation for your content or any other reason, please contact me immediately. I can't promise you won't. But I can promise I'll investigate. I am your advocate in Google and I will continue to act as such as long as I am employed by Google.

My opinion on the 3.5 model is my own and may or may not be that of my employer or Google DeepMind. I posted this and I believe it is true:

The cost bump is the conversion tax from communication prior to agentic state transition prior. Once base models are trained natively for agentic workloads, that tax will have been paid.

Models that are chat tuned are not good agents. I've been working on this research for going on a decade now and been working on the mathematics for more than 25 years. I'm not making an argument from authority, but the "one model to rule them all" approach in my mind has always been dead wrong.

Likewise models that are agent tuned will not be fun to talk to. Language is designed to communicate. Agent tuning has to bend natural language away from meaningful to a reader and toward state transition for a machine. This is no small task.

My hope is that these concepts bear out in real-world use cases and am personally testing 3.5 now in this regard. We'll see. As much as we know about building these models and how they behave at inference time, I cannot fully express how much there is that we do not know. That does not mean we don't have mathematics that matches observed data, we do. It means some structure in very high dimensional space is difficult to understand even without addressing dynamics.

I'll have much more on this as I do the engineering personally and working with our customers to better understand the model behaves in real-world agentic workflows. However, this is the correct direction for models in my opinion. There needs to be a bifurcation of models for chat and models for agentic composition.

Chat continuations have always been a surrogate primitive. We need base models trained to be agents. IMHO that's the future.

Sorry for the rant. Anytime you want to be hosted at the Google office in SF, I'll do so proudly. That won't change.